Develop Reference

Is there a more efficient way to loop through a pandas dataframe? [duplicate]

I have a pandas dataframe, df:
c1 c2
0 10 100
1 11 110
2 12 120
How do I iterate over the rows of this dataframe? For every row, I want to be able to access its elements (values in cells) by the name of the columns. For example:
for row in df.rows:
print(row['c1'], row['c2'])
I found a similar question which suggests using either of these:
for date, row in df.T.iteritems():
for row in df.iterrows():
But I do not understand what the row object is and how I can work with it.

DataFrame.iterrows is a generator which yields both the index and row (as a Series):
import pandas as pd
df = pd.DataFrame({'c1': [10, 11, 12], 'c2': [100, 110, 120]})
df = df.reset_index() # make sure indexes pair with number of rows
for index, row in df.iterrows():
print(row['c1'], row['c2'])
10 100
11 110
12 120

How to iterate over rows in a DataFrame in Pandas
Answer: DON'T*!
Iteration in Pandas is an anti-pattern and is something you should only do when you have exhausted every other option. You should not use any function with "iter" in its name for more than a few thousand rows or you will have to get used to a lot of waiting.
Do you want to print a DataFrame? Use DataFrame.to_string().
Do you want to compute something? In that case, search for methods in this order (list modified from here):
Vectorization
Cython routines
List Comprehensions (vanilla for loop)
DataFrame.apply(): i)  Reductions that can be performed in Cython, ii) Iteration in Python space
DataFrame.itertuples() and iteritems()
DataFrame.iterrows()
iterrows and itertuples (both receiving many votes in answers to this question) should be used in very rare circumstances, such as generating row objects/nametuples for sequential processing, which is really the only thing these functions are useful for.
Appeal to Authority
The documentation page on iteration has a huge red warning box that says:
Iterating through pandas objects is generally slow. In many cases, iterating manually over the rows is not needed [...].
* It's actually a little more complicated than "don't". df.iterrows() is the correct answer to this question, but "vectorize your ops" is the better one. I will concede that there are circumstances where iteration cannot be avoided (for example, some operations where the result depends on the value computed for the previous row). However, it takes some familiarity with the library to know when. If you're not sure whether you need an iterative solution, you probably don't. PS: To know more about my rationale for writing this answer, skip to the very bottom.
Faster than Looping: Vectorization, Cython
A good number of basic operations and computations are "vectorised" by pandas (either through NumPy, or through Cythonized functions). This includes arithmetic, comparisons, (most) reductions, reshaping (such as pivoting), joins, and groupby operations. Look through the documentation on Essential Basic Functionality to find a suitable vectorised method for your problem.
If none exists, feel free to write your own using custom Cython extensions.
Next Best Thing: List Comprehensions*
List comprehensions should be your next port of call if 1) there is no vectorized solution available, 2) performance is important, but not important enough to go through the hassle of cythonizing your code, and 3) you're trying to perform elementwise transformation on your code. There is a good amount of evidence to suggest that list comprehensions are sufficiently fast (and even sometimes faster) for many common Pandas tasks.
The formula is simple,
# Iterating over one column - `f` is some function that processes your data
result = [f(x) for x in df['col']]
# Iterating over two columns, use `zip`
result = [f(x, y) for x, y in zip(df['col1'], df['col2'])]
# Iterating over multiple columns - same data type
result = [f(row[0], ..., row[n]) for row in df[['col1', ...,'coln']].to_numpy()]
# Iterating over multiple columns - differing data type
result = [f(row[0], ..., row[n]) for row in zip(df['col1'], ..., df['coln'])]
If you can encapsulate your business logic into a function, you can use a list comprehension that calls it. You can make arbitrarily complex things work through the simplicity and speed of raw Python code.
Caveats
List comprehensions assume that your data is easy to work with - what that means is your data types are consistent and you don't have NaNs, but this cannot always be guaranteed.
The first one is more obvious, but when dealing with NaNs, prefer in-built pandas methods if they exist (because they have much better corner-case handling logic), or ensure your business logic includes appropriate NaN handling logic.
When dealing with mixed data types you should iterate over zip(df['A'], df['B'], ...) instead of df[['A', 'B']].to_numpy() as the latter implicitly upcasts data to the most common type. As an example if A is numeric and B is string, to_numpy() will cast the entire array to string, which may not be what you want. Fortunately zipping your columns together is the most straightforward workaround to this.
*Your mileage may vary for the reasons outlined in the Caveats section above.
An Obvious Example
Let's demonstrate the difference with a simple example of adding two pandas columns A + B. This is a vectorizable operation, so it will be easy to contrast the performance of the methods discussed above.
Benchmarking code, for your reference. The line at the bottom measures a function written in numpandas, a style of Pandas that mixes heavily with NumPy to squeeze out maximum performance. Writing numpandas code should be avoided unless you know what you're doing. Stick to the API where you can (i.e., prefer vec over vec_numpy).
I should mention, however, that it isn't always this cut and dry. Sometimes the answer to "what is the best method for an operation" is "it depends on your data". My advice is to test out different approaches on your data before settling on one.
My Personal Opinion *
Most of the analyses performed on the various alternatives to the iter family has been through the lens of performance. However, in most situations you will typically be working on a reasonably sized dataset (nothing beyond a few thousand or 100K rows) and performance will come second to simplicity/readability of the solution.
Here is my personal preference when selecting a method to use for a problem.
For the novice:
Vectorization (when possible); apply(); List Comprehensions; itertuples()/iteritems(); iterrows(); Cython
For the more experienced:
Vectorization (when possible); apply(); List Comprehensions; Cython; itertuples()/iteritems(); iterrows()
Vectorization prevails as the most idiomatic method for any problem that can be vectorized. Always seek to vectorize! When in doubt, consult the docs, or look on Stack Overflow for an existing question on your particular task.
I do tend to go on about how bad apply is in a lot of my posts, but I do concede it is easier for a beginner to wrap their head around what it's doing. Additionally, there are quite a few use cases for apply has explained in this post of mine.
Cython ranks lower down on the list because it takes more time and effort to pull off correctly. You will usually never need to write code with pandas that demands this level of performance that even a list comprehension cannot satisfy.
* As with any personal opinion, please take with heaps of salt!
Further Reading
10 Minutes to pandas, and Essential Basic Functionality - Useful links that introduce you to Pandas and its library of vectorized*/cythonized functions.
Enhancing Performance - A primer from the documentation on enhancing standard Pandas operations
Are for-loops in pandas really bad? When should I care? - a detailed write-up by me on list comprehensions and their suitability for various operations (mainly ones involving non-numeric data)
When should I (not) want to use pandas apply() in my code? - apply is slow (but not as slow as the iter* family. There are, however, situations where one can (or should) consider apply as a serious alternative, especially in some GroupBy operations).
* Pandas string methods are "vectorized" in the sense that they are specified on the series but operate on each element. The underlying mechanisms are still iterative, because string operations are inherently hard to vectorize.
Why I Wrote this Answer
A common trend I notice from new users is to ask questions of the form "How can I iterate over my df to do X?". Showing code that calls iterrows() while doing something inside a for loop. Here is why. A new user to the library who has not been introduced to the concept of vectorization will likely envision the code that solves their problem as iterating over their data to do something. Not knowing how to iterate over a DataFrame, the first thing they do is Google it and end up here, at this question. They then see the accepted answer telling them how to, and they close their eyes and run this code without ever first questioning if iteration is the right thing to do.
The aim of this answer is to help new users understand that iteration is not necessarily the solution to every problem, and that better, faster and more idiomatic solutions could exist, and that it is worth investing time in exploring them. I'm not trying to start a war of iteration vs. vectorization, but I want new users to be informed when developing solutions to their problems with this library.

First consider if you really need to iterate over rows in a DataFrame. See this answer for alternatives.
If you still need to iterate over rows, you can use methods below. Note some important caveats which are not mentioned in any of the other answers.
DataFrame.iterrows()
for index, row in df.iterrows():
print(row["c1"], row["c2"])
DataFrame.itertuples()
for row in df.itertuples(index=True, name='Pandas'):
print(row.c1, row.c2)
itertuples() is supposed to be faster than iterrows()
But be aware, according to the docs (pandas 0.24.2 at the moment):
iterrows: dtype might not match from row to row
Because iterrows returns a Series for each row, it does not preserve dtypes across the rows (dtypes are preserved across columns for DataFrames). To preserve dtypes while iterating over the rows, it is better to use itertuples() which returns namedtuples of the values and which is generally much faster than iterrows()
iterrows: Do not modify rows
You should never modify something you are iterating over. This is not guaranteed to work in all cases. Depending on the data types, the iterator returns a copy and not a view, and writing to it will have no effect.
Use DataFrame.apply() instead:
new_df = df.apply(lambda x: x * 2, axis = 1)
itertuples:
The column names will be renamed to positional names if they are invalid Python identifiers, repeated, or start with an underscore. With a large number of columns (>255), regular tuples are returned.
See pandas docs on iteration for more details.

You should use df.iterrows(). Though iterating row-by-row is not especially efficient since Series objects have to be created.

While iterrows() is a good option, sometimes itertuples() can be much faster:
df = pd.DataFrame({'a': randn(1000), 'b': randn(1000),'N': randint(100, 1000, (1000)), 'x': 'x'})
%timeit [row.a * 2 for idx, row in df.iterrows()]
# => 10 loops, best of 3: 50.3 ms per loop
%timeit [row[1] * 2 for row in df.itertuples()]
# => 1000 loops, best of 3: 541 µs per loop

You can use the df.iloc function as follows:
for i in range(0, len(df)):
print(df.iloc[i]['c1'], df.iloc[i]['c2'])

You can also use df.apply() to iterate over rows and access multiple columns for a function.
docs: DataFrame.apply()
def valuation_formula(x, y):
return x * y * 0.5
df['price'] = df.apply(lambda row: valuation_formula(row['x'], row['y']), axis=1)

How to iterate efficiently
If you really have to iterate a Pandas dataframe, you will probably want to avoid using iterrows(). There are different methods and the usual iterrows() is far from being the best. itertuples() can be 100 times faster.
In short:
As a general rule, use df.itertuples(name=None). In particular, when you have a fixed number columns and less than 255 columns. See point (3)
Otherwise, use df.itertuples() except if your columns have special characters such as spaces or '-'. See point (2)
It is possible to use itertuples() even if your dataframe has strange columns by using the last example. See point (4)
Only use iterrows() if you cannot the previous solutions. See point (1)
Different methods to iterate over rows in a Pandas dataframe:
Generate a random dataframe with a million rows and 4 columns:
df = pd.DataFrame(np.random.randint(0, 100, size=(1000000, 4)), columns=list('ABCD'))
print(df)
1) The usual iterrows() is convenient, but damn slow:
start_time = time.clock()
result = 0
for _, row in df.iterrows():
result += max(row['B'], row['C'])
total_elapsed_time = round(time.clock() - start_time, 2)
print("1. Iterrows done in {} seconds, result = {}".format(total_elapsed_time, result))
2) The default itertuples() is already much faster, but it doesn't work with column names such as My Col-Name is very Strange (you should avoid this method if your columns are repeated or if a column name cannot be simply converted to a Python variable name).:
start_time = time.clock()
result = 0
for row in df.itertuples(index=False):
result += max(row.B, row.C)
total_elapsed_time = round(time.clock() - start_time, 2)
print("2. Named Itertuples done in {} seconds, result = {}".format(total_elapsed_time, result))
3) The default itertuples() using name=None is even faster but not really convenient as you have to define a variable per column.
start_time = time.clock()
result = 0
for(_, col1, col2, col3, col4) in df.itertuples(name=None):
result += max(col2, col3)
total_elapsed_time = round(time.clock() - start_time, 2)
print("3. Itertuples done in {} seconds, result = {}".format(total_elapsed_time, result))
4) Finally, the named itertuples() is slower than the previous point, but you do not have to define a variable per column and it works with column names such as My Col-Name is very Strange.
start_time = time.clock()
result = 0
for row in df.itertuples(index=False):
result += max(row[df.columns.get_loc('B')], row[df.columns.get_loc('C')])
total_elapsed_time = round(time.clock() - start_time, 2)
print("4. Polyvalent Itertuples working even with special characters in the column name done in {} seconds, result = {}".format(total_elapsed_time, result))
Output:
A B C D
0 41 63 42 23
1 54 9 24 65
2 15 34 10 9
3 39 94 82 97
4 4 88 79 54
... .. .. .. ..
999995 48 27 4 25
999996 16 51 34 28
999997 1 39 61 14
999998 66 51 27 70
999999 51 53 47 99
[1000000 rows x 4 columns]
1. Iterrows done in 104.96 seconds, result = 66151519
2. Named Itertuples done in 1.26 seconds, result = 66151519
3. Itertuples done in 0.94 seconds, result = 66151519
4. Polyvalent Itertuples working even with special characters in the column name done in 2.94 seconds, result = 66151519
This article is a very interesting comparison between iterrows and itertuples

I was looking for How to iterate on rows and columns and ended here so:
for i, row in df.iterrows():
for j, column in row.iteritems():
print(column)

We have multiple options to do the same, and lots of folks have shared their answers.
I found the below two methods easy and efficient to do:
DataFrame.iterrows()
DataFrame.itertuples()
Example:
import pandas as pd
inp = [{'c1':10, 'c2':100}, {'c1':11,'c2':110}, {'c1':12,'c2':120}]
df = pd.DataFrame(inp)
print (df)
# With the iterrows method
for index, row in df.iterrows():
print(row["c1"], row["c2"])
# With the itertuples method
for row in df.itertuples(index=True, name='Pandas'):
print(row.c1, row.c2)
Note: itertuples() is supposed to be faster than iterrows()

You can write your own iterator that implements namedtuple
from collections import namedtuple
def myiter(d, cols=None):
if cols is None:
v = d.values.tolist()
cols = d.columns.values.tolist()
else:
j = [d.columns.get_loc(c) for c in cols]
v = d.values[:, j].tolist()
n = namedtuple('MyTuple', cols)
for line in iter(v):
yield n(*line)
This is directly comparable to pd.DataFrame.itertuples. I'm aiming at performing the same task with more efficiency.
For the given dataframe with my function:
list(myiter(df))
[MyTuple(c1=10, c2=100), MyTuple(c1=11, c2=110), MyTuple(c1=12, c2=120)]
Or with pd.DataFrame.itertuples:
list(df.itertuples(index=False))
[Pandas(c1=10, c2=100), Pandas(c1=11, c2=110), Pandas(c1=12, c2=120)]
A comprehensive test
We test making all columns available and subsetting the columns.
def iterfullA(d):
return list(myiter(d))
def iterfullB(d):
return list(d.itertuples(index=False))
def itersubA(d):
return list(myiter(d, ['col3', 'col4', 'col5', 'col6', 'col7']))
def itersubB(d):
return list(d[['col3', 'col4', 'col5', 'col6', 'col7']].itertuples(index=False))
res = pd.DataFrame(
index=[10, 30, 100, 300, 1000, 3000, 10000, 30000],
columns='iterfullA iterfullB itersubA itersubB'.split(),
dtype=float
)
for i in res.index:
d = pd.DataFrame(np.random.randint(10, size=(i, 10))).add_prefix('col')
for j in res.columns:
stmt = '{}(d)'.format(j)
setp = 'from __main__ import d, {}'.format(j)
res.at[i, j] = timeit(stmt, setp, number=100)
res.groupby(res.columns.str[4:-1], axis=1).plot(loglog=True);

To loop all rows in a dataframe you can use:
for x in range(len(date_example.index)):
print date_example['Date'].iloc[x]

for ind in df.index:
print df['c1'][ind], df['c2'][ind]

Update: cs95 has updated his answer to include plain numpy vectorization. You can simply refer to his answer.
cs95 shows that Pandas vectorization far outperforms other Pandas methods for computing stuff with dataframes.
I wanted to add that if you first convert the dataframe to a NumPy array and then use vectorization, it's even faster than Pandas dataframe vectorization, (and that includes the time to turn it back into a dataframe series).
If you add the following functions to cs95's benchmark code, this becomes pretty evident:
def np_vectorization(df):
np_arr = df.to_numpy()
return pd.Series(np_arr[:,0] + np_arr[:,1], index=df.index)
def just_np_vectorization(df):
np_arr = df.to_numpy()
return np_arr[:,0] + np_arr[:,1]

Sometimes a useful pattern is:
# Borrowing #KutalmisB df example
df = pd.DataFrame({'col1': [1, 2], 'col2': [0.1, 0.2]}, index=['a', 'b'])
# The to_dict call results in a list of dicts
# where each row_dict is a dictionary with k:v pairs of columns:value for that row
for row_dict in df.to_dict(orient='records'):
print(row_dict)
Which results in:
{'col1':1.0, 'col2':0.1}
{'col1':2.0, 'col2':0.2}

To loop all rows in a dataframe and use values of each row conveniently, namedtuples can be converted to ndarrays. For example:
df = pd.DataFrame({'col1': [1, 2], 'col2': [0.1, 0.2]}, index=['a', 'b'])
Iterating over the rows:
for row in df.itertuples(index=False, name='Pandas'):
print np.asarray(row)
results in:
[ 1. 0.1]
[ 2. 0.2]
Please note that if index=True, the index is added as the first element of the tuple, which may be undesirable for some applications.

In short
Use vectorization if possible
If an operation can't be vectorized - use list comprehensions
If you need a single object representing the entire row - use itertuples
If the above is too slow - try swifter.apply
If it's still too slow - try a Cython routine
Benchmark

There is a way to iterate throw rows while getting a DataFrame in return, and not a Series. I don't see anyone mentioning that you can pass index as a list for the row to be returned as a DataFrame:
for i in range(len(df)):
row = df.iloc[[i]]
Note the usage of double brackets. This returns a DataFrame with a single row.

For both viewing and modifying values, I would use iterrows(). In a for loop and by using tuple unpacking (see the example: i, row), I use the row for only viewing the value and use i with the loc method when I want to modify values. As stated in previous answers, here you should not modify something you are iterating over.
for i, row in df.iterrows():
df_column_A = df.loc[i, 'A']
if df_column_A == 'Old_Value':
df_column_A = 'New_value'
Here the row in the loop is a copy of that row, and not a view of it. Therefore, you should NOT write something like row['A'] = 'New_Value', it will not modify the DataFrame. However, you can use i and loc and specify the DataFrame to do the work.

Sometimes loops really are better than vectorized code
As many answers here correctly point out, your default plan in Pandas should be to write vectorized code (with its implicit loops) rather than attempting an explicit loop yourself. But the question remains whether you should ever write loops in Pandas, and if so what's the best way to loop in those situations.
I believe there is at least one general situation where loops are appropriate: when you need to calculate some function that depends on values in other rows in a somewhat complex manner. In this case, the looping code is often simpler, more readable, and less error prone than vectorized code.
The looping code might even be faster too, as you'll see below, so loops might make sense in cases where speed is of utmost importance. But really, those are just going to be subsets of cases where you probably should have been working in numpy/numba (rather than Pandas) to begin with, because optimized numpy/numba will almost always be faster than Pandas.
Let's show this with an example. Suppose you want to take a cumulative sum of a column, but reset it whenever some other column equals zero:
import pandas as pd
import numpy as np
df = pd.DataFrame( { 'x':[1,2,3,4,5,6], 'y':[1,1,1,0,1,1] } )
# x y desired_result
#0 1 1 1
#1 2 1 3
#2 3 1 6
#3 4 0 4
#4 5 1 9
#5 6 1 15
This is a good example where you could certainly write one line of Pandas to achieve this, although it's not especially readable, especially if you aren't fairly experienced with Pandas already:
df.groupby( (df.y==0).cumsum() )['x'].cumsum()
That's going to be fast enough for most situations, although you could also write faster code by avoiding the groupby, but it will likely be even less readable.
Alternatively, what if we write this as a loop? You could do something like the following with NumPy:
import numba as nb
#nb.jit(nopython=True) # Optional
def custom_sum(x,y):
x_sum = x.copy()
for i in range(1,len(df)):
if y[i] > 0: x_sum[i] = x_sum[i-1] + x[i]
return x_sum
df['desired_result'] = custom_sum( df.x.to_numpy(), df.y.to_numpy() )
Admittedly, there's a bit of overhead there required to convert DataFrame columns to NumPy arrays, but the core piece of code is just one line of code that you could read even if you didn't know anything about Pandas or NumPy:
if y[i] > 0: x_sum[i] = x_sum[i-1] + x[i]
And this code is actually faster than the vectorized code. In some quick tests with 100,000 rows, the above is about 10x faster than the groupby approach. Note that one key to the speed there is numba, which is optional. Without the "#nb.jit" line, the looping code is actually about 10x slower than the groupby approach.
Clearly this example is simple enough that you would likely prefer the one line of pandas to writing a loop with its associated overhead. However, there are more complex versions of this problem for which the readability or speed of the NumPy/numba loop approach likely makes sense.

There are so many ways to iterate over the rows in Pandas dataframe. One very simple and intuitive way is:
df = pd.DataFrame({'A':[1, 2, 3], 'B':[4, 5, 6], 'C':[7, 8, 9]})
print(df)
for i in range(df.shape[0]):
# For printing the second column
print(df.iloc[i, 1])
# For printing more than one columns
print(df.iloc[i, [0, 2]])

The easiest way, use the apply function
def print_row(row):
print row['c1'], row['c2']
df.apply(lambda row: print_row(row), axis=1)

Probably the most elegant solution (but certainly not the most efficient):
for row in df.values:
c2 = row[1]
print(row)
# ...
for c1, c2 in df.values:
# ...
Note that:
the documentation explicitly recommends to use .to_numpy() instead
the produced NumPy array will have a dtype that fits all columns, in the worst case object
there are good reasons not to use a loop in the first place
Still, I think this option should be included here, as a straightforward solution to a (one should think) trivial problem.

You can also do NumPy indexing for even greater speed ups. It's not really iterating but works much better than iteration for certain applications.
subset = row['c1'][0:5]
all = row['c1'][:]
You may also want to cast it to an array. These indexes/selections are supposed to act like NumPy arrays already, but I ran into issues and needed to cast
np.asarray(all)
imgs[:] = cv2.resize(imgs[:], (224,224) ) # Resize every image in an hdf5 file

df.iterrows() returns tuple(a, b) where a is the index and b is the row.

This example uses iloc to isolate each digit in the data frame.
import pandas as pd
a = [1, 2, 3, 4]
b = [5, 6, 7, 8]
mjr = pd.DataFrame({'a':a, 'b':b})
size = mjr.shape
for i in range(size[0]):
for j in range(size[1]):
print(mjr.iloc[i, j])

Disclaimer: Although here are so many answers which recommend not using an iterative (loop) approach (and I mostly agree), I would still see it as a reasonable approach for the following situation:
Extend a dataframe with data from an API
Let's say you have a large dataframe which contains incomplete user data. Now you have to extend this data with additional columns, for example, the user's age and gender.
Both values have to be fetched from a backend API. I'm assuming the API doesn't provide a "batch" endpoint (which would accept multiple user IDs at once). Otherwise, you should rather call the API only once.
The costs (waiting time) for the network request surpass the iteration of the dataframe by far. We're talking about network round trip times of hundreds of milliseconds compared to the negligibly small gains in using alternative approaches to iterations.
One expensive network request for each row
So in this case, I would absolutely prefer using an iterative approach. Although the network request is expensive, it is guaranteed being triggered only once for each row in the dataframe. Here is an example using DataFrame.iterrows:
Example
for index, row in users_df.iterrows():
user_id = row['user_id']
# Trigger expensive network request once for each row
response_dict = backend_api.get(f'/api/user-data/{user_id}')
# Extend dataframe with multiple data from response
users_df.at[index, 'age'] = response_dict.get('age')
users_df.at[index, 'gender'] = response_dict.get('gender')

Some libraries (e.g. a Java interop library that I use) require values to be passed in a row at a time, for example, if streaming data. To replicate the streaming nature, I 'stream' my dataframe values one by one, I wrote the below, which comes in handy from time to time.
class DataFrameReader:
def __init__(self, df):
self._df = df
self._row = None
self._columns = df.columns.tolist()
self.reset()
self.row_index = 0
def __getattr__(self, key):
return self.__getitem__(key)
def read(self) -> bool:
self._row = next(self._iterator, None)
self.row_index += 1
return self._row is not None
def columns(self):
return self._columns
def reset(self) -> None:
self._iterator = self._df.itertuples()
def get_index(self):
return self._row[0]
def index(self):
return self._row[0]
def to_dict(self, columns: List[str] = None):
return self.row(columns=columns)
def tolist(self, cols) -> List[object]:
return [self.__getitem__(c) for c in cols]
def row(self, columns: List[str] = None) -> Dict[str, object]:
cols = set(self._columns if columns is None else columns)
return {c : self.__getitem__(c) for c in self._columns if c in cols}
def __getitem__(self, key) -> object:
# the df index of the row is at index 0
try:
if type(key) is list:
ix = [self._columns.index(key) + 1 for k in key]
else:
ix = self._columns.index(key) + 1
return self._row[ix]
except BaseException as e:
return None
def __next__(self) -> 'DataFrameReader':
if self.read():
return self
else:
raise StopIteration
def __iter__(self) -> 'DataFrameReader':
return self
Which can be used:
for row in DataFrameReader(df):
print(row.my_column_name)
print(row.to_dict())
print(row['my_column_name'])
print(row.tolist())
And preserves the values/ name mapping for the rows being iterated. Obviously, is a lot slower than using apply and Cython as indicated above, but is necessary in some circumstances.

As the accepted answer states, the fastest way to apply a function over rows is to use a vectorized function, the so-called NumPy ufuncs (universal functions).
But what should you do when the function you want to apply isn't already implemented in NumPy?
Well, using the vectorize decorator from numba, you can easily create ufuncs directly in Python like this:
from numba import vectorize, float64
#vectorize([float64(float64)])
def f(x):
#x is your line, do something with it, and return a float
The documentation for this function is here: Creating NumPy universal functions

Along with the great answers in this post I am going to propose Divide and Conquer approach, I am not writing this answer to abolish the other great answers but to fulfill them with another approach which was working efficiently for me. It has two steps of splitting and merging the pandas dataframe:
PROS of Divide and Conquer:
You don't need to use vectorization or any other methods to cast the type of your dataframe into another type
You don't need to Cythonize your code which normally takes extra time from you
Both iterrows() and itertuples() in my case were having the same performance over entire dataframe
Depends on your choice of slicing index, you will be able to exponentially quicken the iteration. The higher index, the quicker your iteration process.
CONS of Divide and Conquer:
You shouldn't have dependency over the iteration process to the same dataframe and different slice. Meaning if you want to read or write from other slice, it maybe difficult to do that.
=================== Divide and Conquer Approach =================
Step 1: Splitting/Slicing
In this step, we are going to divide the iteration over the entire dataframe. Think that you are going to read a CSV file into pandas df then iterate over it. In may case I have 5,000,000 records and I am going to split it into 100,000 records.
NOTE: I need to reiterate as other runtime analysis explained in the other solutions in this page, "number of records" has exponential proportion of "runtime" on search on the df. Based on the benchmark on my data here are the results:
Number of records | Iteration rate [per second]
========================================
100,000 | 500
500,000 | 200
1,000,000 | 50
5,000,000 | 20
Step 2: Merging
This is going to be an easy step, just merge all the written CSV files into one dataframe and write it into a bigger CSV file.
Here is the sample code:
# Step 1 (Splitting/Slicing)
import pandas as pd
df_all = pd.read_csv('C:/KtV.csv')
df_index = 100000
df_len = len(df)
for i in range(df_len // df_index + 1):
lower_bound = i * df_index
higher_bound = min(lower_bound + df_index, df_len)
# Splitting/slicing df (make sure to copy() otherwise it will be a view
df = df_all[lower_bound:higher_bound].copy()
'''
Write your iteration over the sliced df here
using iterrows() or intertuples() or ...
'''
# Writing into CSV files
df.to_csv('C:/KtV_prep_' + str(i) + '.csv')
# Step 2 (Merging)
filename = 'C:/KtV_prep_'
df = (pd.read_csv(f) for f in [filename + str(i) + '.csv' for i in range(ktv_len // ktv_index + 1)])
df_prep_all = pd.concat(df)
df_prep_all.to_csv('C:/KtV_prep_all.csv')
Reference:
Efficient way of iteration over datafreame
Concatenate CSV files into one Pandas Dataframe

Which convention should I adopt for functions that add columns to DataFrames?

Given the following DataFrame,
df = pd.DataFrame({
'a': [3, 5, 7, 9],
'b': [4, 12, 24, 40]
})
I would like to perform a number of steps that each add columns to this DataFrame. To keep things clean, I'd like to provide one function for each column-adding task. I can think of at least three ways to do it. What are the pros and cons of each approach?
Method 1:
Create a Series with apply() and add it to the DataFrame as a new column:
def method1(row):
return np.sqrt(row['a']**2 + row['b']**2)
# This is the line that will appear in main()
df['c'] = df.apply(method1, axis=1)
Pros:
Clear in calling signature that a column is being added
Good for parallelization
Cons:
Does not permit multiple columns to be added by the function
Method 2:
Pass the entire DataFrame into the function
def method2(df):
df['c'] = np.sqrt(df['a']**2 + df['b']**2)
method2(df)
Pros:
Less typing
Avoids the behavior of apply that runs the function twice for the first row (only minor concern)
Cons:
Less amenable to parallelization
"In-place" operation (without explicit request) inconsistent with pandas convention, and the effect of function on df is obscure in main()
Method 3:
Reconstruct DataFrame from lengthened rows with apply()
def method3(row):
row['c'] = np.sqrt(row['a']**2 + row['b']**2)
return row
df = df.apply(method3, axis=1)
Pros:
Permits multiple columns to be added by single function
Good for parallelization
Cons:
The fact that columns are added is not clear in the call signature
Performance (?)

Your misunderstanding here is that apply parallelizes operations... it doesn't.
All it is, is wrapper around a loop. It works something like this:
df['c'] = df.apply(method1, axis=1)
Is equivalent to,
temp = []
for i, row in df.iterrows():
temp.append(method1(row))
df['c'] = temp
On the other hand, just doing
df['c'] = (df['a'] ** 2 + df['b'] ** 2) ** .5
Implicitly vectorizes the entire operation. This is because numpy has many of its basic routines implemented in C, and uses SIMD operations, making this faster than a traditional loop. Time this versus the version with apply and you understand.
df = pd.concat([df] * 100000, ignore_index=True)
%timeit df.apply(method1, axis=1)
1 loop, best of 3: 19.1 s per loop
%timeit (df['a'] ** 2 + df['b'] ** 2) ** .5
100 loops, best of 3: 14.7 ms per loop
The rule of thumb is to always avoid loopy/loop-like solutions, unless your operations are so complex that they cannot be avoided. In that situation, use a simple for loop, the overhead actually works out to be less than apply. You'll thank yourself for it.

My preference in order of priority:
Method 2: This is vectorised and optimal for most datasets. Remember that under the hood pandas uses numpy, so performing vectorised calculations with np.sqrt is efficient.
Method 1: df.apply is not vectorised. It's a thinly veiled loop. Yes, you can optimise the loop with tools such as numba or cython, but then you might as well hold your data and do all your calculations in numpy and produce your pd.DataFrame for purely aesthetic reasons once your calculations are completed.
Method 3: Opaque and unnecessary use of df.apply. You would never choose this method ahead of Method 1, therefore it comes last.

Pandas - Explanation on apply function being slow

Apply function seems to work very slow with a large dataframe (about 1~3 million rows).
I have checked related questions here, like Speed up Pandas apply function, and Counting within pandas apply() function, it seems the best way to speed it up is not to use apply function :)
For my case, I have two kinds of tasks to do with the apply function.
First: apply with lookup dict query
f(p_id, p_dict):
return p_dict[p_dict['ID'] == p_id]['value']
p_dict = DataFrame(...) # it's another dict works like lookup table
df = df.apply(f, args=(p_dict,))
Second: apply with groupby
f(week_id, min_week_num, p_dict):
return p_dict[(week_id - min_week_num < p_dict['WEEK']) & (p_dict['WEEK'] < week_id)].ix[:,2].mean()
f_partial = partial(f, min_week_num=min_week_num, p_dict=p_dict)
df = map(f, df['WEEK'])
I guess for the fist case, it could be done with dataframe join, while I am not sure about resource cost for such join on a large dataset.
My question is:
Is there any way to substitute apply in the two above cases?
Why is apply so slow? For the dict lookup case, I think it should be O(N), it shouldn't cost that much even if N is 1 million.

Concerning your first question, I can't say exactly why this instance is slow. But generally, apply does not take advantage of vectorization. Also, apply returns a new Series or DataFrame object, so with a very large DataFrame, you have considerable IO overhead (I cannot guarantee this is the case 100% of the time since Pandas has loads of internal implementation optimization).
For your first method, I assume you are trying to fill a 'value' column in df using the p_dict as a lookup table. It is about 1000x faster to use pd.merge:
import string, sys
import numpy as np
import pandas as pd
##
# Part 1 - filling a column by a lookup table
##
def f1(col, p_dict):
return [p_dict[p_dict['ID'] == s]['value'].values[0] for s in col]
# Testing
n_size = 1000
np.random.seed(997)
p_dict = pd.DataFrame({'ID': [s for s in string.ascii_uppercase], 'value': np.random.randint(0,n_size, 26)})
df = pd.DataFrame({'p_id': [string.ascii_uppercase[i] for i in np.random.randint(0,26, n_size)]})
# Apply the f1 method as posted
%timeit -n1 -r5 temp = df.apply(f1, args=(p_dict,))
>>> 1 loops, best of 5: 832 ms per loop
# Using merge
np.random.seed(997)
df = pd.DataFrame({'p_id': [string.ascii_uppercase[i] for i in np.random.randint(0,26, n_size)]})
%timeit -n1 -r5 temp = pd.merge(df, p_dict, how='inner', left_on='p_id', right_on='ID', copy=False)
>>> 1000 loops, best of 5: 826 µs per loop
Concerning the second task, we can quickly add a new column to p_dict that calculates a mean where the time window starts at min_week_num and ends at the week for that row in p_dict. This requires that p_dict is sorted by ascending order along the WEEK column. Then you can use pd.merge again.
I am assuming that min_week_num is 0 in the following example. But you could easily modify rolling_growing_mean to take a different value. The rolling_growing_mean method will run in O(n) since it conducts a fixed number of operations per iteration.
n_size = 1000
np.random.seed(997)
p_dict = pd.DataFrame({'WEEK': range(52), 'value': np.random.randint(0, 1000, 52)})
df = pd.DataFrame({'WEEK': np.random.randint(0, 52, n_size)})
def rolling_growing_mean(values):
out = np.empty(len(values))
out[0] = values[0]
# Time window for taking mean grows each step
for i, v in enumerate(values[1:]):
out[i+1] = np.true_divide(out[i]*(i+1) + v, i+2)
return out
p_dict['Means'] = rolling_growing_mean(p_dict['value'])
df_merged = pd.merge(df, p_dict, how='inner', left_on='WEEK', right_on='WEEK')

Search for elements by timestamp in a sorted pandas dataframe

I have a very large pandas dataframe/series with millions of elements.
And I need to find all the elements for which timestamp is < than t0.
So normally what I would do is:
selected_df = df[df.index < t0]
This takes ages. As I understand when pandas searches it goes through every element of the dataframe. However I know that my dataframe is sorted hence I can break the loop as soon as the timestamp is > t0. I assume pandas doesn't know that dataframe is sorted and searches through all timestamps.
I have tried to use pandas.Series instead - still very slow.
I have tried to write my own loop like:
boudery = 0
ticks_time_list = df.index
tsearch = ticks_time_list[0]
while tsearch < t0:
tsearch = ticks_time_list[boudery]
boudery += 1
selected_df = df[:boudery]
This takes even longer than pandas search.
The only solution I can see atm is to use Cython.
Any ideas how this can be sorted without C involved?

It doesn't really seem to take ages for me, even with a long frame:
>>> df = pd.DataFrame({"A": 2, "B": 3}, index=pd.date_range("2001-01-01", freq="1 min", periods=10**7))
>>> len(df)
10000000
>>> %timeit df[df.index < "2001-09-01"]
100 loops, best of 3: 18.5 ms per loop
But if we're really trying to squeeze out every drop of performance, we can use the searchsorted method after dropping down to numpy:
>>> %timeit df.iloc[:df.index.values.searchsorted(np.datetime64("2001-09-01"))]
10000 loops, best of 3: 51.9 µs per loop
>>> df[df.index < "2001-09-01"].equals(df.iloc[:df.index.values.searchsorted(np.datetime64("2001-09-01"))])
True
which is many times faster.

(I'm not very familiar with Pandas, but this describes a very generic idea - you should be able to apply it. If necessary, adapt the Pandas-specific functions.)
You could try to use a more efficient search. At the moment you are using a linear search, going through all the elements. Instead, try this
ticks_time_list=df.index
tsearch_min = 0
tsearch_max = len(ticks_time_list)-1 #I'm not sure on whether this works on a pandas dataset
while True:
tsearch_middle = int((tsearch_max-tsearch_min)/2)
if ticks_time_list[tsearch_middle] < t0:
tsearch_min = tsearch_middle
else:
tsearch_max = tsearch_middle
if tsearch_max == tsearch_min:
break
# tsearch_max == tsearch_min and is the value of the index you are looking for
Instead of opening every single element, and looking at the time stamp, it instead tries to find the "boundary" by always narrowing down the search space by cutting it into half.

Fast pandas filtering

I want to filter a pandas dataframe, if the name column entry has an item in a given list.
Here we have a DataFrame
x = DataFrame(
[['sam', 328], ['ruby', 3213], ['jon', 121]],
columns=['name', 'score'])
Now lets say we have a list, ['sam', 'ruby'] and we want to find all rows where the name is in the list, then sum the score.
The solution I have is as follows:
total = 0
names = ['sam', 'ruby']
for name in names:
identified = x[x['name'] == name]
total = total + sum(identified['score'])
However when the dataframe gets extremely large, and the list of names gets very large too, everything is very very slow.
Is there any faster alternative?
Thanks

Try using isin (thanks to DSM for suggesting loc over ix here):
In [78]: x = pd.DataFrame([['sam',328],['ruby',3213],['jon',121]], columns = ['name', 'score'])
In [79]: names = ['sam', 'ruby']
In [80]: x['name'].isin(names)
Out[80]:
0 True
1 True
2 False
Name: name, dtype: bool
In [81]: x.loc[x['name'].isin(names), 'score'].sum()
Out[81]: 3541
CT Zhu suggests a faster alternative using np.in1d:
In [105]: y = pd.concat([x]*1000)
In [109]: %timeit y.loc[y['name'].isin(names), 'score'].sum()
1000 loops, best of 3: 413 µs per loop
In [110]: %timeit y.loc[np.in1d(y['name'], names), 'score'].sum()
1000 loops, best of 3: 335 µs per loop

If I need to search on a field, I have noticed that it helps immensely if I change the index of the DataFrame to the search field. For one of my search and lookup requirements I got a performance improvement of around 500%.
So in your case the following could be used to search and filter by name.
df = pd.DataFrame([['sam', 328], ['ruby', 3213], ['jon', 121]],
columns=['name', 'score'])
names = ['sam', 'ruby']
df_searchable = df.set_index('name')
df_searchable[df_searchable.index.isin(names)]
Update Dec-21
Updates are driven by the comments on this answer.
Looking at the details of my use case, its not magic that is happening here. My use case was that of running millions of look-ups on a column with around 45k values. From what I remember, it was a lookup on US zip-codes. Understandably, once the set_index has incurred it's one time optimization cost, subsequent look-ups become way faster. The overall effect is magnified because of the large number of look-ups, the cost of optimization getting amortized over all the numerous look-ups.
The impressive performance improvement number is essentially due to the highly amortized optimization cost.

If your data repeats a lot of values, try using the 'categorical' data type for that column and then applying boolean filtering. Much more flexible than using indices and, at least in my case, much faster.
data = pd.read_csv('data.csv', dtype={'name':'category'})
data[(data.name=='sam')&(data.score>1)]
or
names=['sam','ruby']
data[data.name.isin(names)]
For the ~15 million row, ~200k unique terms dataset I'm working with in pandas 1.2, %timeit results are:
boolean filter on object column: 608ms
.loc filter on same object column as index: 281ms
boolean filter on same object column as 'categorical' type: 16ms
From there, add the .sum() or whatever aggregation function you're looking for.